[R-meta] publication bias on BLUPs
Hi Yefeng, As far as I know, this is a wide open methodological question. My entirely speculative first guess would be to use something like a Vevea-Hedges selection model (e.g., metafor::selmodel()) to estimate the mean and variance of the effect size distribution, and then calculate the BLUPs as precision-weighted averages of the study-specific point estimates and the overall average effect estimate. But like I said, this is nothing more than speculation, with no supporting theory or evidence that it's valid. You could perhaps do something similar with regression-based corrections or with trim-and-fill, but those methods involve very rough approximations that aren't based on generative models for the data. I'm therefore less keen on using them for this sort of bias correction. James On Tue, Jul 4, 2023 at 11:21?PM Yefeng Yang via R-sig-meta-analysis <
r-sig-meta-analysis at r-project.org> wrote:
Hi experts,
I am writing to you with utmost gratitude for the incredible support I
received during my recent query about BLUPs on this platform. The
explanations provided by experts like James, Wolfgang, and others were
truly invaluable, and their selfless dedication to helping countless
individuals like myself without any financial support is truly commendable.
Over the past few weeks, my fascination with BLUPs has only grown. Some
knowledge might be fairly easy for experts, but they are new to me. Now I
come across two new questions about BLUPs in the context of meta-analysis.
My questions are two-fold: (1) whether there is a way to test whether
publication bias has an impact on the estimates of BLUPs, (2) if yes, how
to get the bias-adjusted BLUPs.
Let's use the dataset in metafor as an example:
# calculate es and var
dat <- escalc(measure="OR", ai=ai, n1i=n1i, ci=ci, n2i=n2i,
data=dat.egger2001, subset=-16)
# fit RE model
res <- rma(yi, vi, data=dat)
# use the Egger test to test publication bias, assuming funnel asymmetry
is the proxy of publication bias
regtest(res, model="lm") # we see there is a correlation between mean and
se
# we use a correction method. There are quite a few, such as
precision-based method (i.e., intercept from Egger's test) and
trim-and-fill . Here we use trim-and-fill
trimfill(res) # we get the bias-adjusted effect
So how to answer my two questions? If no readily available solution, any
comments or inspiration? Very much appreciated.
Best,
Yefeng
[[alternative HTML version deleted]]
_______________________________________________ R-sig-meta-analysis mailing list @ R-sig-meta-analysis at r-project.org To manage your subscription to this mailing list, go to: https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis